This is a tricky question. If the SDM has a moderate dynamic range, the Vref can be generated on chip. Using a bandgap reference to drive a fully-differential amplifier will give Vhigh and Vlow, and therefore the Vref. The design of the fully-differential amplifier with minimum power requires a bit of knowledge. There is a document from Analog Device talking about this.
If your SDM has a very high SNDR, say >100dB, then Vref is good to be off-chip. Just my two cents.